> OK, hadn't run into that personally, but let's fix that too. Anything > that can be run with "installcheck" has to satisfy the restrictions > of being re-runnable and careful about what global names it uses.
Actually, now that I re-read that thread, maybe what we need to do is mark test_pg_dump as something not to be run by "make installcheck"? I'm not quite sure what *does* run it, but the context is evidently that it's supposed to be run by some overarching script that's then going to run pg_dump on the ending database state. So it's not meant to be run against a generic pre-existing installation, and it doesn't have to follow the rules for being safe for that --- but then we need to be sure that it doesn't get called that way if someone does "make installcheck" in a parent directory.
On my phone atm, but this feels very deja vu...
Isn’t this the one run from pg_upgrade’s tests? We don’t want to break that (and hopefully we haven’t but maybe something did...). Pretty sure we had nearly the same discussion this past fall...